Slow crash dump speed on ARM server

My Ampere eMAG server takes about two hours to save a crash dump after panic, a write speed of about 3 MB/s to an NVMe SSD that is capable of hundreds of times the performance if not a thousand times.  Is this normal?  Does a crash dump write tiny blocks using synchronous writes or something like that?

/var/crash/info.last has
  Blocksize: 512
  Compression: none

The blocksize is suspicious.  Is the kernel writing 512 bytes at a time to an SSD with native 4K blocks?  That would be bad.

Is there a sysctl parameter I can set to make writes faster?  I did want the dump this time but this performance would not be acceptable for a production server.

(Math: "Dumping 16023 out of 130745 MB" took ~6000 seconds.  A 20 GB dump took over 2 hours.)